Document Topic Modeling Framework

A comprehensive framework for unsupervised topic modeling and document classification using transformer-based embeddings. This project includes both the research notebooks for model development and a complete interactive web application.

Project Overview

This project combines modern natural language processing techniques to discover latent topics in text documents without manual labeling:

• Transformer Embeddings: Leverage RoBERTa to generate high-quality document representations
• Dimensionality Reduction: Apply UMAP to preserve semantic relationships while reducing dimensions
• Clustering: Implement K-means clustering to discover natural topic groups
• Interactive Visualization: Explore document-topic relationships through interactive visualizations

The implementation is provided both as research notebooks and as a production-ready web application.

Repository Structure

└─── Notebooks/
│    │    Advanced_Topic_Modeling.ipynb
│    │    Transformers_Embedding_Benchmark.ipynb
└─── application
│    └─── static/
│    └─── templates/
│    └─── topic_model_export/
│    │    app.py
│    │    requirements.txt
│    README.md

Research Notebooks

Advanced_Topic_Modeling.ipynb

This main notebook implements a comprehensive framework for unsupervised topic modeling using transformer-based embeddings, tested on the BBC News Dataset (2,225 documents across 5 categories):

1. Data Understanding & Preprocessing: Thorough exploration of the BBC News dataset including category distribution, document length analysis, and word frequency analysis.
2. Embedding Pipeline: Implementation of RoBERTa embeddings for document representation, with methods to handle document length constraints.
3. Dimensionality Reduction: Comparison of PCA, t-SNE, and UMAP techniques, with UMAP providing the best performance.
4. Clustering Implementation: K-means and hierarchical clustering algorithms with evaluation metrics like ARI, homogeneity, and silhouette scores.
5. Topic Extraction & Labeling: Methods to automatically extract distinctive terms and generate meaningful topic labels for each cluster.
6. Visualization Tools: Creation of various visualizations including topic proximity maps, document-topic distributions, and interactive visualizations.

Transformers_Embedding_Benchmark.ipynb

This notebook focuses on benchmarking different transformer models and embedding techniques to identify the optimal configuration:

1. Model Comparison: Evaluation of 5 transformer models (BERT, RoBERTa, DistilBERT, MPNet, and ALBERT) with the same BBC News dataset.
2. Embedding Method Assessment: Comparison of two embedding approaches - CLS token (using the classification token) and Mean token (averaging all token embeddings).
3. Performance Metrics: Thorough evaluation using clustering quality metrics (ARI, Homogeneity, V-measure, Silhouette score) and processing efficiency.
4. Results Analysis: Detailed visualizations and comparisons of model performance, with recommendations for the optimal configuration.
5. Recommendations: Clear conclusions about which model and embedding method perform best for topic modeling tasks.

Web Application

The /application folder contains a production-ready Flask web application that implements the topic modeling framework with an intuitive user interface.

Application Overview

This application uses advanced natural language processing techniques to automatically classify text documents into topics without requiring any manual labeling. The system combines several state-of-the-art machine learning components:

Technical Architecture

• RoBERTa Embeddings: A state-of-the-art transformer model that converts text into high-dimensional semantic representations (768 dimensions)
• UMAP Dimensionality Reduction: Reduces the high-dimensional embeddings to a manageable space while preserving important semantic relationships
• K-means Clustering: Groups documents with similar meanings into distinct topic

Classification Process

1. Document Embedding: When you input a document, the system generates a RoBERTa embedding that captures its semantic meaning
2. Dimensional Projection: This embedding is then projected into the same space as pre-trained topic clusters using UMAP
3. Proximity Analysis: The system calculates the proximity to each topic cluster and assigns confidence scores
4. Topic Assignment: The document is classified into the most likely topic, with confidence scores shown for all categories
5. Key Term Extraction: Distinctive terms from the document are identified and displayed to explain classification reasoning

Features

• Text Input Options: Paste text directly or select from example documents
• Real-time Classification: Process documents instantly and see results without page refresh
• Confidence Analysis: View probability distribution across all potential topics
• Document-specific Key Terms: Extract and highlight distinctive terms from the input document
• Interactive Visualization: Explore how the document positions in relation to topic clusters in 2D space
• Responsive Design: Works seamlessly on desktop and mobile devices

Screenshots

Getting Started

Prerequisites

• Python 3.7+
• Required libraries listed in requirements.txt

Installation

1. Clone the repository:

git clone https://github.com/imanerh/Topic-Modeling.git
cd Topic-Modeling

2. Install required packages:

pip install -r requirements.txt

Running the Web Application

1. Navigate to the application directory:

cd application

2. Run the Flask app:

python app.py

3. Open your browser and go to:

http://localhost:5000

Results

The framework achieves excellent topic separation on the BBC News dataset:

• High Clustering Metrics: 0.85+ homogeneity and V-measure scores
• Strong Category Alignment: 90%+ alignment with ground truth categories
• Interpretable Topics: Clear and distinctive terms characterizing each topic